AI BriefingNace Releases Weight-Space Quantized GGUFs for Qwen 9B, 27B, 35B, and Flash-Next
·2026.11.14 04:23
Key point
The release uses an LLM to natively edit weights, outperforming Unsloth and EXL3 in KL divergence and top-1 accuracy at 128K context.
Details
- Nace released GGUF weights for Qwen 9B, 27B, 35B, and Flash-Next using a Weight-Space Programming (WSP) method where an LLM agent proposes and edits tensor precisions to minimize KL divergence.
- The quants achieve the highest top-1 accuracy among files of the same size or smaller: 9B Q3_K_XL reaches 90.5% at 4.7 GB, 35B Q3_K_S reaches 91.2% at 15.3 GB, 27B Q4_K_XL reaches 97.7% at 17.3 GB, and Flash-Next IQ4_XS reaches 93.0% at 93 GB.
- Long-context stability is a key differentiator; at 128K tokens, Nace files maintain lower KL divergence than competitors, whereas standard quants often degrade significantly from 512-token performance.
- MMLU benchmark results show that Q3-Q4 quantization is nearly indistinguishable from higher precision on knowledge-heavy tasks, with gains mostly within noise margins.
- The method relies on an agentic swarm (Astra/Fable/Carbon) to hillclimb optimization, outperforming brute-force approaches and existing SOTA quants like Unsloth and EXL3 in specific precision classes.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
DetLLM Achieves Bit-Identical LLM Inference Across Hardware Using Integer-Only Arithmetic
Unsloth Releases Dynamic V3.1 Quantizations for Qwen3.8-27B and New 1-bit Data Types for Qwen3.8-2.4TNew Arm NEON kernel speeds up llama.cpp prompt processing by up to 6.5xFigma's unreleased editing model, post-trained on Qwen 3.8 27B, beats top third-party model on internal evals